Research on super-resolution image reconstruction method based on multi-axis attention

Abstract

Aiming at the problems of poor target quality and insufficient information mining in existing small-sample data images, a multi-axis attention-based super-resolution image reconstruction algorithm (MT-SRGAN) is proposed. Firstly, the generator adopts a combination of Transformer and convolutional neural network as the main architecture to focus on global information and extract deep and shallow features. Secondly, the discriminator uses dilated convolution multi-axis attention residual blocks as the main architecture to alleviate gradient vanishing and perform detailed feature extraction on images. Finally, global average pooling is used to reduce the complexity of model parameters. This method optimizes both the generator and the discriminator simultaneously. Tests are conducted on MobileNetV2 and Densenet121 classification networks using standard datasets RSOD and NWPU. The results show that this method has significant improvements compared with other current methods.